Google CloudVertex AI WorkbenchCompute

`google_notebooks_instance` cost estimation

A managed JupyterLab notebook VM for data science. Priced at the underlying Compute Engine rate per hour, plus boot disk and any attached GPUs.

A google_notebooks_instance is a managed Vertex AI Workbench notebook, a JupyterLab environment running on a Compute Engine VM. Vertex AI adds no markup over the VM, so the cost is the standard Compute Engine machine-type rate for as long as the instance runs, plus its boot disk and any GPUs.

The machine type sets the hourly rate (an e2-standard-4 is roughly $0.134/hour, ~$98/month if left running). A persistent disk for the notebook environment bills per GB-month. If you attach a GPU for training or inference, that adds the GPU's hourly rate, which often dwarfs the CPU cost. There's no scale-to-zero, the instance bills whenever it's running, though you can stop it manually.

c3x reads the machine_type and prices the instance-hours at 730 hours/month, adding boot disk and GPU where configured. The classic waste here is leaving a notebook (especially a GPU notebook) running overnight and on weekends. Stopping idle notebooks, or using idle-shutdown, is the main lever; the estimate makes the running cost visible.

Terraform example

A minimal but realistic configuration that C3X can estimate.

resource "google_notebooks_instance" "ds" {
  name         = "research-notebook"
  location     = "us-central1-a"
  machine_type = "e2-standard-4"

  vm_image {
    project      = "deeplearning-platform-release"
    image_family = "common-cpu-notebooks"
  }
}

Pricing dimensions

What you actually pay for when you provision google_notebooks_instance.

Dimension	Unit	What's being charged
Instance hours	per hour	Standard Compute Engine rate for the machine_type while the notebook runs. c3x assumes 730 hours/month unless stopped. $0.134/hour for e2-standard-4
Boot disk	per GB-month	Persistent disk for the notebook environment. Billed even while the instance is stopped. $0.10/GB-month (balanced PD)
GPU (optional)	per GPU-hour	Attached accelerators (T4, L4, A100) add their hourly rate, often far exceeding the CPU cost. Only when configured.

Sample C3X output

Example output from c3x estimate (e2-standard-4, CPU only):

google_notebooks_instance.ds
└─ Instance hours (e2-standard-4)   730  hours    $97.84

OVERALL TOTAL                                     $97.84
(boot disk & any GPU billed on top)

Optimization tips

Common ways to reduce google_notebooks_instance cost without changing the workload.

Enable idle shutdown

All idle running time

Workbench can auto-stop after a period of inactivity. Enabling idle shutdown stops the instance (halting compute charges) when a researcher walks away, the single biggest saving for notebooks.

Stop GPU notebooks when not training

GPU idle hours (large)

A GPU attached to a notebook bills its full hourly rate whenever the instance runs, even while you're writing code, not training. Stop the instance between training sessions, or detach the GPU.

Right-size the machine type

Proportional to size

Notebooks are often provisioned larger than the analysis needs. For exploration and light work, a smaller machine type runs fine and costs proportionally less.

FAQ

How does c3x estimate Vertex AI Workbench cost?

It reads machine_type and prices instance-hours at the standard Compute Engine rate (no Vertex markup) for 730 hours/month, plus boot disk and any attached GPU. Stopping the instance reduces real cost.

Does Vertex AI add a markup over the VM price?

No. A Workbench instance pays the standard Compute Engine machine-type rate. The managed JupyterLab environment is included at no extra per-hour charge.

Why do notebooks waste money?

They're left running, especially GPU notebooks, when no one's using them. Compute bills whenever the instance runs. Idle-shutdown and manually stopping notebooks between sessions are the main savings.

How much does adding a GPU cost?

A lot. A T4 adds ~$0.35/hour, larger accelerators much more, on top of the CPU rate, and it bills whenever the instance runs (not just during training). Stop GPU notebooks when idle.

Is the boot disk charged when stopped?

Yes. The persistent boot disk bills per GB-month whether the instance is running or stopped. Only compute (and GPU) charges stop when you stop the instance.